mini-batch training HAN with DGL

This is an attempt to implement mini-batch training with HAN, which used RandomWalkSample to get sub-graphs. Original code from dgl HAN

Dependency:

dgl 0.6.1
ogb 1.3.0

---

mini-batch training on ACMRaw(DGL dataset):

python train_sample.py --dataset ACMRaw --num_neighbors 20 --batch_size 32 --hidden_units 8

It can match the full graph training result: dgl HAN

EarlyStopping counter: 10 out of 10
Epoch 23 | Val loss 0.5857 | Val Accuracy 0.9177 | Val Micro f1 0.9177 | Val Macro f1 0.9176
Test loss 0.3041 | Test Accuracy 0.9176 | Test Micro f1 0.9176 | Test Macro f1 0.9174

---

mini-batch training on ogbn-mag:

python train_sample.py --dataset ogbn-mag --num_neighbors 100 --batch_size 1024 --hidden_units 512

the result is bad (T T):

Epoch 43 | loss: 3.4052 | train_acc: 0.2359 | train_micro_f1: 0.2359 | train_macro_f1: 0.0625
EarlyStopping counter: 10 out of 10
Val loss 3.4083 | Val Accuracy 0.2482 | Val Micro f1 0.2482 | Val Macro f1 0.0388
Test loss 3.2047 | Test Accuracy 0.2691 | Test Micro f1 0.2691 | Test Macro f1 0.0383

---

full training on PAP based meta-graph(need more than 110G memory):

python train_PAPMAG_full.py

result not good:

Epoch 522 | loss: 3.3740 | train_acc: 0.2356 | train_micro_f1: 0.2356 | train_macro_f1: 0.0411
EarlyStopping counter: 100 out of 100
Val loss 3.1961 | Val Accuracy 0.2638 | Val Micro f1 0.2638 | Val Macro f1 0.0382
Test loss 3.1105 | Test Accuracy 0.2914 | Test Micro f1 0.2914 | Test Macro f1 0.0425